We currently make our space-filling designs on-the-fly using functions in DiceDesign. They use random numbers and don't always produce the best results.
This website has compiled designs that have been optimized for varying numbers of parameters and sizes. They are free of charge and I've already scraped them.
I'd like to:
For grid_latin_hypercube(), use the Audze-Eglais designs whenever possible (if they are available for the chosen number of parameters and size). Otherwise, do what we've done before.
Add a new grid_max_min() function to look up the L2 and L1 designs. We might consider moving the default grid in tune to use this. It goes up to 10 parameters so we are probably fine.
This is also made me feel that we should adjust the grid_max_entropy() function to raise the values of the radius and the number of iterations to make them more stable. We could also add a tries argument that will use different random numbers and take the design with the best results.
We currently make our space-filling designs on-the-fly using functions in
DiceDesign
. They use random numbers and don't always produce the best results.This website has compiled designs that have been optimized for varying numbers of parameters and sizes. They are free of charge and I've already scraped them.
I'd like to:
grid_latin_hypercube()
, use the Audze-Eglais designs whenever possible (if they are available for the chosen number of parameters and size). Otherwise, do what we've done before.grid_max_min()
function to look up the L2 and L1 designs. We might consider moving the default grid in tune to use this. It goes up to 10 parameters so we are probably fine.This is also made me feel that we should adjust the
grid_max_entropy()
function to raise the values of the radius and the number of iterations to make them more stable. We could also add atries
argument that will use different random numbers and take the design with the best results.